{"id":"b247ecf7-c86e-4fc8-a4f6-6b018b07d781","arxiv_id":"2507.04035","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"New divergence-kernel formulas compute scores of SDEs with multiplicative noise via pathwise Monte Carlo, without hyperbolicity assumptions.","lead":"This paper derives formulas for computing the score, the gradient of log probability density, of stochastic dynamical systems driven by multiplicative noise. The formulas are turned into a pathwise Monte Carlo algorithm and demonstrated on a 40-dimensional Lorenz-96 system.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1.1 is only a formal limit: the backward-adapted SDE for ν and the injectivity/invertibility assumptions are not established, so the continuous-time score formula lacks proof despite plausible numerics.","rationale":"The reader's weakest assumption correctly identifies the formal continuous-time passage as a central concern. My stress test supports that concern and adds two unstated conditions: the discrete theorem presumes injectivity of g_b, which fails on positive-probability events for some multiplicative-noise Euler maps, and the backward-adapted integral in the ν-SDE is never rigorously defined beyond the discrete sum. I checked an exactly solvable case (dx=x dB) analytically and the divergence-kernel formula reproduces the correct score, so I do not see evidence that the claim is false; it is unproved as stated. The numerical experiments use only constant α, so they do not exercise the backward-adapted aspect, and the Lorenz 96 check is an integrated observable without error bars. These are addressable gaps, so the reader's CONDITIONAL verdict remains appropriate.","tokens_in":18473,"tokens_out":22627,"duration_ms":255935,"concrete_test":"Take a one-dimensional SDE with a known score (e.g., dx=−x^3dt+(0.5+e^{−x^2})dB) and compute E[ν_T|x_T] from Theorem 1.1 with a genuinely non-constant backward-adapted schedule, such as α_t=1+(W_T−W_t), at several time steps Δt=10^{-4}, 5×10^{-5}, 2.5×10^{-5}, comparing against a spectral Fokker-Planck reference. If the bias does not vanish as Δt→0, the formal limit or the backward-integral definition fails. Separately, run the same check with constant α to confirm that the σ^{-1} scaling in the dB term is implemented as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 1.1) is obtained by formally passing the discrete-time recurrence (Theorem 4.6) to the continuous limit. This passage is not proved. Section 2 concedes 'all integrations, averages, and change of limits are legit', and Equation (10) is the only definition of the backward-adapted stochastic integral appearing in the ν-SDE. Three concrete gaps remain. First, the diffusion coefficient −(∇σν^T + ∇^2σ + ασ^{-1})dB contains a backward-adapted integrand, an anticipative integral whose existence, uniqueness, and discretization independence are not established. Second, the limit drops O(ΔBΔt), O(Δt^2), and higher-order terms without the L^2 estimates needed to justify the identity E[ν_T|x_T]. Third, the discrete Theorem 4.6 inherits the injectivity assumption on g_b made in Section 3.2; for Euler maps with multiplicative noise, for example σ(x)=1+0.5 sin x, the map g_b(x)=x+F(x)Δt+σ(x)b is non-injective for large |b|, which has positive probability, so even the 'rigorous' discrete statement needs an unstated assumption or a sum over preimages. Thus the main continuous-time claim is not established beyond formal manipulation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a 'divergence-kernel method' for computing the score function ∇ log h_T of random dynamical systems. For discrete-time systems x_{n+1}=f(x_n)+σ(x_n)b_n with i.i.d. noise b_n, it derives kernel, divergence, and mixed formulas expressing the score as a conditional expectation of a recursively defined covector process ν_n (Theorems 4.2, 4.3, 4.6, 4.7). It then formally passes to the continuous-time limit d x_t = F(x_t) dt + σ(x_t) dB_t and states formal SDE versions, including an arbitrary backward-adapted schedule α_t (Theorem 1.1) and a version avoiding ∇ log h_0 (Theorem 5.5). Numerical experiments on a 1D OU process and a 40-dimensional Lorenz 96 system support the formulas qualitatively, and the code is provided.","tokens_in":18786,"tokens_out":6835,"duration_ms":66972,"significance":"If the continuous-time formula were proven, it would give a pathwise unbiased estimator for scores of SDEs with multiplicative noise, with no hyperbolicity assumption, which is directly relevant to score-based generative modeling and linear response theory. The discrete-time results provide a clean conceptual framework that mixes divergence and kernel differentiation methods. The paper is transparent about the formal nature of the continuous-time limit and posts reproducible code; the zero-mean identity check in Section 6.2 is a useful sanity test. However, the central Theorem 1.1 is explicitly formal, and the discrete-time theorems require an injectivity assumption that is not stated in their hypotheses.","major_comments":[{"comment":"The continuous-time result is not established. Section 2 disclaims technical assumptions ('We just assume that all integrations, averages, and change of limits are legit'), and the proof of Theorem 5.1 explicitly says 'Neglect the sum of O(ΔBΔt) terms' without providing L^2 estimates or a convergence argument. Moreover, the ν-SDE in Theorem 1.1 contains the term −(∇σν^T + ∇^2σ + ασ^{−1})dB with backward-adapted α; the stochastic integral is anticipative, and Equation (10) gives only a formal limit, not a definition with existence and uniqueness. Consequently Theorem 1.1 is a conjecture, not a theorem, and the main claim of the paper is unsupported. This must be addressed, either by a rigorous proof of the limit or by clearly reframing the paper's contribution as the rigorous discrete-time results plus a conjectured continuous-time limit.","section":"Section 5 and Theorem 1.1"},{"comment":"The derivation of the divergence formulas assumes injectivity of the map g_b(x) = f(x) + σ(x)b. This assumption is mentioned only parenthetically after Equation (3) ('if it is n-to-1, then we need to further sum over its preimage'), but it is not included in the statements of Lemmas 3.2 and 4.4 or Theorems 4.3, 4.6, and 4.7. For the Euler discretization g_b(x) = x + F(x)Δt + σ(x)b used in Section 5, injectivity can fail with positive probability when σ is non-monotone, e.g., σ(x) = 1 + 0.5 sin x and |b| large. Thus the 'rigorous' discrete-time theorems need either an explicit injectivity assumption, a small-Δt condition, or a version with sums over preimages.","section":"Section 3.2 and Theorems 4.3, 4.6, 4.7"},{"comment":"The numerical evidence is too weak to verify the exactness of the continuous-time formula. In Section 6.1 the score is compared with the empirical density only visually and averaged over 10 paths per bin, and in Section 6.2 the check is the zero-mean identity E[Φ(x_T)ν_T·v_T] = −1, which tests the integrated score rather than the pointwise conditional score. A more convincing test would use a system with a known closed-form score (e.g., a linear SDE) and report convergence as Δt → 0 with statistical error bars; the small O(ΔBΔt) terms dropped in Section 5 could produce a bias undetectable by the current checks.","section":"Section 6"}],"minor_comments":[{"comment":"The notation σ is used both for the diffusion coefficient and for the σ-algebra generated by (x_0, x_1); this is confusing, particularly in Section 3.3.","section":"Section 2"},{"comment":"The proof of Theorem 4.6 says 'expand everything and check terms match Lemma 4.5' and does not show the expansion; please include the details.","section":"Theorem 4.6"},{"comment":"In Theorem 4.7, α_n = 1/n is undefined at n = 0; the v_0 term vanishes, but this should be stated explicitly.","section":"Theorem 4.7"},{"comment":"Equation (9) treats ΔB as both a vector and a covector in ∇ log k(ΔB) = −ΔB/Δt; please clarify the convention.","section":"Section 5.1, Eq. (9)"},{"comment":"There are several typos: 'fomula' in Theorem 1.1, 'independ' in Section 4.1, and an agreement error in the abstract ('formulas which does not involve').","section":"Throughout"},{"comment":"The caption states that the score is averaged on 10 paths; this is too few to indicate statistical reliability, and the caption should state that the plot is illustrative.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is refreshingly honest about the formal nature of the continuous-time limit, but for a mathematics probability journal with MSC primary 60H07/60J60, an unproven central theorem is a serious issue. The discrete-time core is a useful contribution; the paper could be resubmitted after either a rigorous treatment of the continuous-time limit or a reframing as a discrete-time paper with the SDE results stated as conjectures. The self-citation pattern is extensive but contextual."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this paper derives a pathwise Monte Carlo estimator for scores of SDEs with multiplicative noise, and as far as I can tell the core identity is new and works. The discrete-time theorems (4.6, 4.7) are proven cleanly, and the paper is honest that the continuous-time Theorem 1.1 is a formal limit. That honesty matters—it tells you what you can trust and what still needs work.\n\nThe genuinely new bits are the recursive divergence-kernel combination, the back-ward adapted schedule, and the version without the initial score (Theorem 5.5). The one-step identity is an algebraic recombination, but the multi-step recursion and its continuous limit are not in the cited literature. The numerical tests, while qualitative and without error bars, do demonstrate the method on a 40-dimensional multiplicative-noise system, which is more than most formal papers in this area. The self-citations are contextual; the earlier papers genuinely build up the framework.\n\nSoft spots, in order of seriousness. First, the continuous-time theorem is formal: the paper says so explicitly, and the passage from Theorem 4.6 to the SDE drops O(ΔBΔt) terms without L2 estimates. The backward-adapted stochastic integral defining the ν SDE is not given a rigorous meaning. Second, even the \"rigorous\" discrete theorems inherit an injectivity assumption on g_b that the paper states but does not check. For Euler maps with multiplicative noise, g_b can fail to be injective with positive probability, so the discrete statement also needs a preimage sum or a restriction to a set where injectivity holds. This is fixable, but worth flagging. Third, the numerics are supportive but not quantitative—no error bars, no convergence study, just a histogram and slope comparisons. They suggest the formula is right, not that it is proven.\n\nMy bottom line: this is a serious paper with a real idea, written by someone who knows the literature and is upfront about the gaps. The formal continuous theorem is the kind of thing that should be published with a clear \"rigorous proof later\" marker, because the discrete version already gives a practical algorithm. I would not desk-reject it. I would send it to a referee who knows Malliavin calculus and Monte Carlo methods, with the expectation that the formal status and the injectivity point be handled in revision or explicitly scoped.\n\nFor you: if you work on score-based generative models or linear response for diffusions, this is worth reading and citing. I'd bring it to reading group.","headline":"A genuinely useful formal algorithm with rigorous discrete-time underpinnings, but the headline continuous-time theorem is a formal limit and the numerics are qualitative.","tokens_in":19222,"tokens_out":1231,"would_cite":true,"duration_ms":14486,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H07","60J60","65D25","65C30","65C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper derives the divergence-kernel formula for scores of random systems: for an SDE with multiplicative noise, the terminal score equals the conditional expectation of a covector process driven by the same noise, yielding a pathwise…","keywords":["Score function","Divergence method","Kernel-differentiation method","Cameron-Martin-Girsanov","Diffusion process","Multiplicative noise","Monte Carlo estimator","Random dynamical systems"],"falsifier":"Simulate many paths of a one-dimensional multiplicative-noise SDE, for example $dx = -x^3\\,dt + (0.5+e^{-x^2})\\,dB$, with $T=3$ and small $\\Delta t$; compute the conditional average $\\mathbb{E}[\\nu_T \\mid x_T]$ from Theorem 1.1 and compare it with a finite-difference or kernel-density estimate of $\\nabla\\log h_T$ at several endpoints, using enough samples that Monte Carlo error is negligible. A systematic discrepancy that does not vanish as $\\Delta t \\to 0$ would disprove the central formula.","tokens_in":18308,"feed_emoji":"🎲","tokens_out":8869,"duration_ms":95284,"temperature":0.7,"pith_summary":"The paper tries to establish that the score of the terminal density of a random dynamical system can be represented as the conditional expectation of a single covector process evolved along the same random path. The representation mixes the divergence method with the kernel-differentiation method, and it is designed to work where each alone fails: multiplicative noise, non-hyperbolic dynamics, and long time horizons. The discrete-time version is proved rigorously; the continuous-time SDE version is derived by a formal limit that the paper states as an assumption. If the formula holds, scores can be estimated by pathwise Monte Carlo simulation, which would give diffusion models and linear-response computations a tool they currently lack.","feed_headline":"Pathwise estimator computes scores for multiplicative-noise SDEs","feed_subtitle":"Formula needs no hyperbolicity and handles singular initial conditions, tested on a 40-dimensional system","key_machinery":"The load-bearing object is the one-step divergence-kernel identity: for $x_1 = f(x_0)+\\sigma(x_0)b_0$, the score is expressed as a weighted combination of the kernel derivative $\\nabla\\log k(b_0)/\\sigma(x_0)$ and the divergence term $g^{-1}_{b_0*}(\\nabla\\log h_0 - \\operatorname{div}g_{b_0*})$, with weights controlled by a matrix $\\alpha_1$. Recursing this identity yields the forward covector process in Theorem 4.6, which is rigorous in discrete time. The continuous-time limit replaces $g_{b_*}^{-1}$ and $\\operatorname{div}g_{b_*}$ with the approximations in Lemmas 5.2 and 5.3 and uses $\\nabla\\log k(\\Delta B) = -\\Delta B/\\Delta t$ for Gaussian increments; the backward adapted schedule $\\alpha$ is what lets the limit keep the damping term while cancelling the otherwise explosive growth.","core_discovery":"The central claim is Theorem 1.1. For the SDE $$dx_t = F(x_t)\\,dt + \\$\\sigma$(x_t)\\,dB_t,$$ let $h_t$ be the density of $x_t$ and let $\\nu_t$ be the forward covector process solving $$d\\nu = \\big((\\nabla\\$\\sigma$\\nabla\\$\\sigma$^T - \\nabla F^T - \\$\\alpha$)\\nu - \\nabla\\operatorname{div}F + \\$nabla^{2}$\\$\\sigma$\\nabla\\$\\sigma$ + \\nabla\\$\\sigma$\\,\\$\\Delta$\\$\\sigma$\\big)dt - \\big(\\nabla\\$\\sigma$\\,\\nu^T + \\$nabla^{2}$\\$\\sigma$ + \\$\\alpha$\\$sigma^{{-1}}$\\big)dB$$ with initial condition $\\nu_0 = \\nabla\\log h_0(x_0)$. Then $$\\nabla\\log h_T(x_T) = \\mathbb{E}[\\nu_T \\mid x_T].$$ Here $\\alpha_t$ is any backward adapted process, typically nonnegative, that damps exponential growth of $\\nu$ by shifting part of the divergence onto the probability kernel. The same framework gives pure kernel formulas for additive noise, pure divergence formulas for short time, and a schedule that removes $\\nabla\\log h_0$ entirely, so singular initial distributions can be handled.","pith_inferences":["If the formal limit can be made rigorous under mild regularity, the estimator is unbiased for any finite $T$, making it a drop-in score estimator for score-based diffusion models that currently rely on approximations of the initial score.","Because $\\alpha_t$ is arbitrary and only affects variance, adaptive schedules that increase $\\alpha$ in strongly contracting regions and decrease it where noise is small should improve sample efficiency; the paper tests only constant $\\alpha$, so this is a direct testable extension.","The same divergence-kernel split should transfer to parameter derivatives, since the divergence term already carries derivatives of terminal densities; the author's stated next step is linear response for diffusions.","The rigorous discrete-time Theorem 4.6 already gives a valid gradient estimator for discrete-time generative models, so the method does not have to wait for a fully rigorous continuous-time proof to be useful."],"forward_implications":["For any finite horizon, the score of an SDE with multiplicative noise can be estimated by simulating the state and covector SDEs together and averaging over paths conditioned on the same endpoint, with no hyperbolicity assumption.","Choosing $\\alpha$ large enough, for example constant $\\alpha > \\frac1T\\log|D_T|$, prevents the covector $\\nu$ from growing exponentially with $T$, so the estimator does not inherit the divergence method's long-time blow-up.","The schedule $\\alpha_t = 1/t$ removes $\\nabla\\log h_0$ from the expression entirely, so the formula applies to singular initial distributions where the initial score is not defined.","For additive noise the formula reduces to a pure kernel expression, and for short time to a pure divergence expression, unifying the two classical methods.","The numerical demonstration on a 40-dimensional chaotic system with multiplicative noise and a singular initial distribution shows the estimator is implementable in high dimensions."],"supporting_citations":[{"why":"Supplies the recursive divergence formula and the notation $\\operatorname{div}g_* = \\nabla|g_*|/|g_*|$ that the divergence-kernel derivation extends.","marker":"[35]"},{"why":"Gives the matrix-notation derivation of the divergence formula and the covector algebra used in the proofs.","marker":"[30]"},{"why":"Cameron-Martin-Girsanov theorem is the basis of the kernel-differentiation part of the formula.","marker":"[6]"},{"why":"Standard stochastic-analysis reference for the Girsanov and kernel view of path densities used throughout the paper.","marker":"[27]"},{"why":"The path-kernel method for $dB$-type perturbations that this paper builds on to handle diffusion-coefficient dependence.","marker":"[33]"},{"why":"Existing score formula for additive-noise SDEs that the paper compares against and improves.","marker":"[28]"},{"why":"Lays out the triad program of path, divergence, and kernel methods and motivates mixing divergence with kernel.","marker":"[32]"},{"why":"Provides the 40-dimensional chaotic test model used in the numerical demonstration.","marker":"[25]"}],"fun_headline_variants":["SDE score formula needs no hyperbolicity","Pathwise scores for multiplicative-noise SDEs","Divergence-kernel scores without hyperbolicity","New formula computes SDE scores pathwise","Score estimator for multiplicative-noise SDEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The continuous-time formula rests on the assumption that the small-time-step limit and all interchanges of integrations, averages, and limits are legitimate; the paper states this assumption explicitly rather than proving it.","fun_headline_variants_meta":{"raw":{"variants":["SDE score formula needs no hyperbolicity","Pathwise scores for multiplicative-noise SDEs","Divergence-kernel scores without hyperbolicity","New formula computes SDE scores pathwise","Score estimator for multiplicative-noise SDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1595,"prompt_tokens":897,"completion_tokens":698,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":628}},"tokens_in":513,"tokens_out":698,"duration_ms":7562,"temperature":1.0,"reasoning_tokens":628,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:56:37.598606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate many paths of a one-dimensional multiplicative-noise SDE, for example $dx = -x^3\\,dt + (0.5+e^{-x^2})\\,dB$, with $T=3$ and small $\\Delta t$; compute the conditional average $\\mathbb{E}[\\nu_T \\mid x_T]$ from Theorem 1.1 and compare it with a finite-difference or kernel-density estimate of $\\nabla\\log h_T$ at several endpoints, using enough samples that Monte Carlo error is negligible. A systematic discrepancy that does not vanish as $\\Delta t \\to 0$ would disprove the central formula.","supporting_citations":[{"cited_title":"Ni and Y","cited_arxiv_id":null,"evidence_quote":"Supplies the recursive divergence formula and the notation $\\operatorname{div}g_* = \\nabla|g_*|/|g_*|$ that the divergence-kernel derivation extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the matrix-notation derivation of the divergence formula and the covector algebra used in the proofs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cameron-Martin-Girsanov theorem is the basis of the kernel-differentiation part of the formula."},{"cited_title":"Malliavin","cited_arxiv_id":null,"evidence_quote":"Standard stochastic-analysis reference for the Girsanov and kernel view of path densities used throughout the paper."},{"cited_title":"Path-Kernel Method for Differentiating Unstable Diffusions","cited_arxiv_id":"2503.00718","evidence_quote":"The path-kernel method for $dB$-type perturbations that this paper builds on to handle diffusion-coefficient dependence."},{"cited_title":"Mirafzali, U","cited_arxiv_id":null,"evidence_quote":"Existing score formula for additive-noise SDEs that the paper compares against and improves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the 40-dimensional chaotic test model used in the numerical demonstration."}],"review_version":1}